In order to solve the problems of image edge blurring，low accuracy caused by insufficient multi-scale feature extraction and large existing models in the current sinter mixture particle size detection methods，a particle size detection model for sinter mixture based on HGM-UNet network is proposed. The model adopts Ghost convolution，which effectively reduces computational complexity and memory consumption by introducing lightweight convolutional kernels and sparse feature graph representation. The UNet encoder part is improved by the Haar wavelet downsampling module，which fully extracts multi-scale features and improves the accuracy of the model. The MSCA attention mechanism is introduced to improve the accuracy of boundary recognition and structural details. The boundary-weighted cross-entropy loss function is added to better learn and retain the boundary information of the target object，and improve the segmentation accuracy of the particle size of the sinter mixture. The sinter mixture data set is established and verified by comparison. The results show that the proposed model is 2. 50% higher than the mIoU，2. 58% higher PA，2. 55% higher Dice coefficient，and 25. 4% lower parameter quantity than the traditional UNet network.